{
  "model": "dE_pct(N) = A - S * (N/Nstar)/(1 + N/Nstar)",
  "s_cap": {
    "NF4": 75.0,
    "INT8": 50.0
  },
  "arch_label": {
    "turing": "Turing (T4)",
    "ada": "Ada (RTX 4090/4090D)",
    "blackwell": "Blackwell (RTX 5090)",
    "ampere": "Ampere (A100/A800)",
    "hopper": "Hopper (H100/H200)"
  },
  "borrow": {
    "hopper": "ampere"
  },
  "curves": {
    "ada": {
      "NF4": {
        "A": 106.0936,
        "S": 90.55,
        "Nstar": 0.381,
        "resid_std": 3.6947,
        "n_min": 0.5,
        "n_max": 3.0,
        "crossover_b": null,
        "anchors": [
          {
            "N": 0.5,
            "dE": 56.09,
            "model": "Qwen2-0.5B",
            "gpu": "RTX 4090D"
          },
          {
            "N": 1.1,
            "dE": 33.35,
            "model": "TinyLlama-1.1B",
            "gpu": "RTX 4090D"
          },
          {
            "N": 1.5,
            "dE": 38.6,
            "model": "Qwen2-1.5B",
            "gpu": "RTX 4090D"
          },
          {
            "N": 3.0,
            "dE": 25.22,
            "model": "Qwen2.5-3B",
            "gpu": "RTX 4090D"
          }
        ],
        "loo_mae": 8.724
      }
    },
    "ampere": {
      "INT8": {
        "A": 180.8252,
        "S": 127.9389,
        "Nstar": 10.082,
        "resid_std": 2.4297,
        "n_min": 7.0,
        "n_max": 14.0,
        "crossover_b": null,
        "anchors": [
          {
            "N": 7.0,
            "dE": 130.83,
            "model": "Mistral-7B",
            "gpu": "A800"
          },
          {
            "N": 9.0,
            "dE": 117.18,
            "model": "Yi-1.5-9B",
            "gpu": "A800"
          },
          {
            "N": 14.0,
            "dE": 107.41,
            "model": "Qwen2.5-14B",
            "gpu": "A800"
          }
        ],
        "loo_mae": null
      },
      "NF4": {
        "A": -1.0293,
        "S": 0.0,
        "Nstar": 200.0,
        "resid_std": 2.7146,
        "n_min": 7.0,
        "n_max": 14.0,
        "crossover_b": null,
        "anchors": [
          {
            "N": 7.0,
            "dE": -4.09,
            "model": "Mistral-7B",
            "gpu": "A800"
          },
          {
            "N": 9.0,
            "dE": -1.51,
            "model": "Yi-1.5-9B",
            "gpu": "A800"
          },
          {
            "N": 14.0,
            "dE": 2.51,
            "model": "Qwen2.5-14B",
            "gpu": "A800"
          }
        ],
        "loo_mae": null
      }
    },
    "blackwell": {
      "NF4": {
        "A": 45.8272,
        "S": 104.4224,
        "Nstar": 6.0749,
        "resid_std": 2.8022,
        "n_min": 1.1,
        "n_max": 7.0,
        "crossover_b": 4.751,
        "anchors": [
          {
            "N": 1.1,
            "dE": 26.49,
            "model": "TinyLlama-1.1B",
            "gpu": "RTX 5090"
          },
          {
            "N": 1.5,
            "dE": 29.42,
            "model": "Qwen2-1.5B",
            "gpu": "RTX 5090"
          },
          {
            "N": 3.0,
            "dE": 11.74,
            "model": "Qwen2.5-3B",
            "gpu": "RTX 5090"
          },
          {
            "N": 7.0,
            "dE": -11.45,
            "model": "Qwen2-7B",
            "gpu": "RTX 5090"
          }
        ],
        "loo_mae": 4.846
      }
    },
    "turing": {
      "NF4": {
        "A": 7.925,
        "S": 79.5646,
        "Nstar": 19.2186,
        "resid_std": 1.316,
        "n_min": 1.1,
        "n_max": 7.0,
        "crossover_b": 2.126,
        "anchors": [
          {
            "N": 1.1,
            "dE": 4.56,
            "model": "TinyLlama-1.1B",
            "gpu": "T4"
          },
          {
            "N": 1.5,
            "dE": 0.22,
            "model": "Qwen2-1.5B",
            "gpu": "T4"
          },
          {
            "N": 3.0,
            "dE": -1.38,
            "model": "Qwen2.5-3B",
            "gpu": "T4"
          },
          {
            "N": 7.0,
            "dE": -13.75,
            "model": "Qwen2-7B",
            "gpu": "T4"
          }
        ],
        "loo_mae": 5.262
      }
    }
  },
  "fp16_energy": {
    "ada": {
      "n_min": 0.5,
      "n_max": 3.0,
      "anchors": [
        {
          "N": 0.5,
          "e_j1k": 1474.16,
          "model": "Qwen2-0.5B",
          "gpu": "RTX 4090D"
        },
        {
          "N": 1.1,
          "e_j1k": 1600.58,
          "model": "TinyLlama-1.1B",
          "gpu": "RTX 4090D"
        },
        {
          "N": 1.5,
          "e_j1k": 2238.87,
          "model": "Qwen2-1.5B",
          "gpu": "RTX 4090D"
        },
        {
          "N": 3.0,
          "e_j1k": 2989.22,
          "model": "Qwen2.5-3B",
          "gpu": "RTX 4090D"
        }
      ]
    },
    "blackwell": {
      "n_min": 1.1,
      "n_max": 7.0,
      "anchors": [
        {
          "N": 1.1,
          "e_j1k": 1659.0,
          "model": "TinyLlama-1.1B",
          "gpu": "RTX 5090"
        },
        {
          "N": 1.5,
          "e_j1k": 2411.09,
          "model": "Qwen2-1.5B",
          "gpu": "RTX 5090"
        },
        {
          "N": 3.0,
          "e_j1k": 3382.64,
          "model": "Qwen2.5-3B",
          "gpu": "RTX 5090"
        },
        {
          "N": 7.0,
          "e_j1k": 5508.56,
          "model": "Qwen2-7B",
          "gpu": "RTX 5090"
        }
      ]
    },
    "turing": {
      "n_min": 1.1,
      "n_max": 7.0,
      "anchors": [
        {
          "N": 1.1,
          "e_j1k": 4251.21,
          "model": "TinyLlama-1.1B",
          "gpu": "T4"
        },
        {
          "N": 1.5,
          "e_j1k": 5731.8,
          "model": "Qwen2-1.5B",
          "gpu": "T4"
        },
        {
          "N": 3.0,
          "e_j1k": 11267.69,
          "model": "Qwen2.5-3B",
          "gpu": "T4"
        },
        {
          "N": 7.0,
          "e_j1k": 21722.65,
          "model": "Qwen2-7B",
          "gpu": "T4"
        }
      ]
    },
    "ampere": {
      "n_min": 7.0,
      "n_max": 14.0,
      "anchors": [
        {
          "N": 7.0,
          "e_j1k": 4402.43,
          "model": "Mistral-7B",
          "gpu": "A800"
        },
        {
          "N": 9.0,
          "e_j1k": 5445.12,
          "model": "Yi-1.5-9B",
          "gpu": "A800"
        },
        {
          "N": 14.0,
          "e_j1k": 7359.98,
          "model": "Qwen2.5-14B",
          "gpu": "A800"
        }
      ]
    }
  },
  "hardware": {
    "turing": {
      "gpu": "T4",
      "mem_bw_gbps": 320.0,
      "fp16_tflops": 65.0,
      "source": "NVIDIA T4 datasheet"
    },
    "ada": {
      "gpu": "RTX 4090D",
      "mem_bw_gbps": 1008.0,
      "fp16_tflops": 330.0,
      "source": "NVIDIA Ada / RTX 4090 datasheet"
    },
    "blackwell": {
      "gpu": "RTX 5090",
      "mem_bw_gbps": 1792.0,
      "fp16_tflops": 419.0,
      "source": "NVIDIA RTX 5090 datasheet"
    },
    "ampere": {
      "gpu": "A800",
      "mem_bw_gbps": 2039.0,
      "fp16_tflops": 312.0,
      "source": "NVIDIA A100/A800 80GB datasheet"
    },
    "hopper": {
      "gpu": "H100",
      "mem_bw_gbps": 3350.0,
      "fp16_tflops": 990.0,
      "source": "NVIDIA H100 SXM datasheet"
    }
  }
}